Car detection is not a trivial task, especially if you want to perform it on ARM devices. Before using the following cascades read carefully this page to get the best performance and to know the terms of usage.

*NEWS*: since June 2016 vision-ary project joined ARGO Vision, an innovative firm that excels in visual recognition. For info about the cascades, please send a message via contact form here.

For any other info about Computer Vision and Artificial Intelligence please contact ARGO Vision.

Full frontal/rear (with partial profiles) car detection cascades, trained with:

  • approx. 7.200 positive samples (randomly sampled).
  • approx 1.3B of negative sub-regions containing outdoor and indoor samples (90%-10%).
  • Training size w=32 h=22 (the best one tested).

LBP: (contact us)

  • Training time: ~3 days.
  • TP: ~ 97.01% of positive training set.
  • FN: ~ 02.99% of positive training set.
  • FP: ~ 0.89e-006% of negative training set.

HAAR: (contact us)

  • Features set: 328384 features.
  • Training time:  ~6 days.
  • TP: ~ 97.02% of positive training set.
  • FN: ~ 02.98% of positive training set.
  • FP: ~ 0.9e-006% of negative training set.

More here:


  1. Dear Sir,

    Hello, I’m a master course student in UST, Korea.
    I’m studing computer vision for object tracking, so I’m interested in your xml files to test my tracking algorithm which I have studied. I already have tested my code with your released xml and realy thank you.
    However, I want to test my code in various detection environment. So, I would appreciate it if you could send me LBP xml file of Car and HoG xml file of Pedestrian to test code with LBP Car and HoG Pedestrian Detection. If you send me xml files, these are helpful for me to study computer vision.
    I am really looking forward to your answer and xml files.

    Whui Kim

    • Dear Kim,
      thanks for contacting us.

      I’ll reply you about the new cascades soon by email, what about the test you did? Any comment? Please share your experience to improve this cascade and offer better xml models.


      • Dear sir,

        Whenever I check my email account, I didn’t get your email. So, I can’t test your cascade xml model. Could you resend your files to my another email account I’m sorry some trouble. Thank you for your help.

        Whui Kim

        • Dear Kim,

          for what concerning the LBP xml file of Car and HoG xml file of Pedestrian two similar cascades are inside two distinct commercial projects, I’m not planning to share them for free until I will close the projects.

          As noticed in the terms of usage, if you use my cascade in research project please cite this site as source. No commercial usage is allowed without our written permission.


    • Dear Akshayakumar ,

      for what concerning the LBP xml file of Car is now inside a commercial project, we are not planning to share it for free soon.

      As noticed in the terms of usage, if you use the cascade in research project please cite this site as source. No commercial usage is allowed without our written permission.


  2. I tried by taking face detection example from OpenCV 2.2 and changed xml file. No detection is happening for me. Am I missing something here ?? I changed min and max object size and scale values.It didnt help

    • Hello Akshayakumar,
      it sounds weird but it’s possible. I tried the classifier with the OpenCV 2.4.2 and it works good. From version to version in OpenCV there are some differences in training and in detection code, one of this difference makes the usage of this cascade possible only with version < 2.4.3. I didn't try an OpenCV older than the 2.3.x, maybe you find another problem with older version. Give a chance to OpenCV 2.4.2 or contact us by email to arrange a train with OpenCV 2.2.x

    • Hello Xou,
      there are two options:
      – converting our cascade from <2.4.3 to 3.1, but it could be not trivial.
      - re-train the cascade from the scratch using the same parameters and datasets.

      What is your favourite one?

      Vision-ary team

      • Hello Vision-ary team,
        I had the same question about use this cascade with OpenCV 3.1.0, what would the process be in order to convert the cascade?
        Thank you.

        • Hello Rkap,
          thanks for contacting us!

          The conversion of the cascade depends from the training process and weights adjustment in the 2.3 version. This process is not trivial, send us a message via contact to understand how to proceed (please specify your goal and your company). Another valid option is to re-train the cascade with the 3.1 version.

          Vision-ary team

  3. Hi!

    Learning about computer vision in my spare time, just wondering do you really use 1.3 “BILLION” samples? If so how large are all those samples and how do you find those samples? Thanks, your site is really informative!


    • Hello Sean,
      thank you!

      We collect images since 2006, our academic period. Please note many samples coming from a single image thanks to classic image transformation (scale, rotation, cropping, etc.). Anyway the Web is full of free datasets (for academic research purposes), or you can extract Gb of images from your DVDs or from your hard disk. Remember always to respect the copyright claims or to spend some dollars to help the academic research buying the dataset for commercial purposes.

      Vision-ary team.

  4. Hi!
    I want to test the detection in a opencv 3.0 environment.Will you upload a latest version of cars_and_truck_cascade_web_HAAR?

    • Hello,

      at the moment it’s not planned this update. If you need the cascade for 3.0 please contact us by the contact form and specify the name of your company and your purposes with the cascades.

      Vision-ary team.

Lascia un commento

Il tuo indirizzo email non sarà pubblicato.